Merge pull request #121 from cog-imperial/add_activations_onnx

Add activations onnx

src/omlt/io/onnx_parser.py

@@ -12,7 +12,7 @@
 )
 from omlt.neuralnet.network_definition import NetworkDefinition
 
-_ACTIVATION_OP_TYPES = ["Relu", "Sigmoid", "LogSoftmax"]
+_ACTIVATION_OP_TYPES = ["Relu", "Sigmoid", "LogSoftmax", "Tanh", "Softplus"]
 _POOLING_OP_TYPES = ["MaxPool"]
 
 
@@ -232,12 +232,12 @@ def _consume_gemm_dense_nodes(self, node, next_nodes):
         attr = _collect_attributes(node)
         alpha = attr["alpha"]
         beta = attr["beta"]
-        assert attr["transB"] == 1
         [in_0, in_1, in_2] = list(node.input)
         input_layer, transformer = self._node_input_and_transformer(in_0)
         weights = self._initializers[in_1]
         # transpose B
-        weights = np.transpose(weights)
+        if attr["transB"] == 1:
+            weights = np.transpose(weights)
         biases = self._initializers[in_2]
 
         input_output_size = _get_input_output_size(input_layer, transformer)

tests/io/test_onnx_parser.py

@@ -56,6 +56,22 @@ def test_gemm(datadir):
     assert layers[3].activation == "logsoftmax"
 
 
+@pytest.mark.skipif(not onnx_available, reason="Need ONNX for this test")
+def test_gemm_transB(datadir):
+    model = onnx.load(datadir.file("gemm_not_transB.onnx"))
+    model_transB = onnx.load(datadir.file("gemm_transB.onnx"))
+    net = load_onnx_neural_network(model)
+    net_transB = load_onnx_neural_network(model_transB)
+    layers = list(net.layers)
+    layers_transB = list(net_transB.layers)
+    assert len(layers) == len(layers_transB)
+    assert layers[1].weights.shape == layers_transB[1].weights.shape
+    assert abs(layers[1].weights[0][0] - layers_transB[1].weights[0][0]) < 1e-05
+    assert abs(layers[1].weights[0][1] - layers_transB[1].weights[1][0]) < 1e-05
+    assert abs(layers[1].weights[1][0] - layers_transB[1].weights[0][1]) < 1e-05
+    assert abs(layers[1].weights[1][1] - layers_transB[1].weights[1][1]) < 1e-05
+
+
 @pytest.mark.skipif(not onnx_available, reason="Need ONNX for this test")
 def test_conv(datadir):
     model = onnx.load(datadir.file("convx1_gemmx1.onnx"))

tests/neuralnet/test_nn_formulation.py

@@ -219,8 +219,10 @@ def test_invalid_layer_type():
 
 def _maxpool_conv_network(inputs):
     input_size = [1, 8, 6]
-    input_bounds = np.empty(input_size, dtype="i,i")
-    input_bounds.fill((-10, 10))
+    input_bounds = {}
+    for i in range(input_size[1]):
+        for j in range(input_size[2]):
+            input_bounds[(0, i, j)] = (-10.0, 10.0)
     net = NetworkDefinition(scaled_input_bounds=input_bounds)
 
     input_layer = InputLayer(input_size)